Transform of X_baseline in qNEHVI (qNoisyExpectedHypervolumeImprovement) and qLNEHVI (qLogNoisyExpectedHypervolumeImprovement) instantiating necessary?

[MoBurmeister]

Hello together,

as mentioned in the title, I was wondering if it is necessary to transform the handed over X_baseline, when initiating the qNEHVI or qLNEHVI acquisition function.

In this tutorial it is done: https://botorch.org/tutorials/multi_objective_bo

But in most of the discussed topics and shared code here in the discussion forum it is not done. Also when sharing part of my code last time (#2470) it was not pointed out/mentioned in the responses.

The docstrings of both acquisition functions do not really provide clarity:

qNEHVI & qLNEHVI:

"X_baseline: A r x d-dim Tensor of r design points that have already
been observed. These points are considered as potential approximate
pareto-optimal design points."

I tried to find out more in the code of botorch, but wasnt really able to get a hint about it.

If it should be done, the transforms of the model should be applied?
And also if transforms are necessary, this can maybe be added to the docstring for more clarity.

Thanks in advance for any help :)

[saitcakmak]

Hi @MoBurmeister. X_baseline should be consistent with the inputs provided to the model. X_baseline is typically same as train_X, so if you pre-process train_X before initializing the model, you should apply the same process to X_baseline as well. This is the path the tutorial takes (which is a bit outdated and can be error-prone). If you're simply using input transforms (e.g., Normalize) within the model, you don't need to do anything to X_baseline.

[bmaxdk]

Transforming X_baseline is necessary if you manually transformed train_X. If BoTorch handles transformations internally, you don’t need to transform X_baseline. Always ensure consistency between X_baseline and the data format expected by your model.